Back-Translation Still Matters: Leveraging Monolingual Data for Low-Resource Machine Translation with LLMs
Back-translation outperformed continued pretraining when low-resource monolingual text was available.
The paper tests LLM translation with small parallel sets plus moderate monolingual data. On Odia and Swahili, back-translation beat continued pretraining by 7.7 to 16.4 XCOMET-XXL points in-domain. With beam-decoded intermediate translations, it also beat a setup using 10 times more parallel data. Across five low-resource languages, back-translation improved translation into the low-resource language by 16% to 173% over parallel-only fine-tuning, with blind human evaluation on Odia and Sanskrit matching the automatic results. ArXiv · AI/CL/LG's note
The paper tests LLM translation with small parallel sets plus moderate monolingual data. On Odia and Swahili, back-translation beat continued pretraining by 7.7 to 16.4 XCOMET-XXL points in-domain. With beam-decoded intermediate translations, it also beat a setup using 10 times more parallel data. Across five low-resource languages, back-translation improved translation into the low-resource language by 16% to 173% over parallel-only fine-tuning, with blind human evaluation on Odia and Sanskrit matching the automatic results. ArXiv · AI/CL/LG's note
score 4